I've been forecasting progress for all the "neolabs", AGI startups that have raised billions and hope to compete with the frontier labs. Thinking Machines Lab, startup Mira Murati founded in February 2025 after leaving the OpenAI CTO job, is the most legible in terms of actually shipping, so it's the most forecastable.
Thinking Machines raised a $2 billion seed at a $12 billion valuation that July, the largest seed round on record at the time. It behaves most like a business compared to any of the other AGI startups. It shipped Tinker, a fine-tuning service people actually use. Last month it shipped Inkling, now the leading US open-weights model. Of my set of neolab forecasts it has the largest compute anchor of the six labs, but their commitment to actually making money seems to reduce the chance that they will actually make it to the AI frontier.
So in this piece, forecasting more specific Thinking Machines outcomes, I focused on the business questions. One thing I noticed immediately is that public reporting on this is really bad, e.g. whether the $50 billion round ever closed. Bloomberg reported the talks, AIwire reported them collapsing after three co-founders left for OpenAI, and nothing authoritative has landed since. No credible revenue estimate exists at all, and the pages that rank for "thinking machines revenue" are auto-generated junk.
So I hope that these forecasts constitute a much more accurate, sourced set of estimates for Thinking Machines' near-term outcomes.
| Question | Forecast |
|---|---|
| Annualized revenue run rate, end of 2027 | $168M ($24M to $850M) |
| A round at $50B+ valuation closes before July 2027 | 24% |
| Any TML model in the Artificial Analysis top 10 before 2028 | 12% |
My forecasts are anchored to this timeline of events:
- February 2025. Murati founds TML with Schulman, Weng, and roughly two dozen researchers from OpenAI and DeepMind.
- July 2025. The $2 billion seed closes at about $12 billion, led by Andreessen Horowitz, with Nvidia, AMD, Accel, Cisco, and Jane Street participating, plus a $10 million check from the government of Albania, Murati's home country.
- October 2025. Tinker launches as a managed fine-tuning platform; it goes generally available in December.
- November 2025. Bloomberg reports funding talks at a $50 billion valuation.
- January 2026. Three co-founders return to OpenAI. AIwire reports the $50B talks collapsed in the aftermath.
- April 2026. A multi-billion dollar Google Cloud expansion, on top of a gigawatt-scale Nvidia Vera Rubin commitment.
- July 2026. Inkling ships, a 975B-parameter open-weights model under Apache 2.0, debuting thirteenth on the Artificial Analysis index and first among US open-weights models. The same month, Lilian Weng leaves for OpenAI.
- August 2026. Secondary trades on Forge mark the company around $13.5 billion. No new primary round has printed.
The revenue median lands at $168M annualized by the end of 2027, with an 80% interval from $24M to $850M. TML's open-weights strategy caps its own upside. Inkling is Apache 2.0, and its hosted inference mostly runs on partners like Together and Fireworks, so TML keeps a hosting cut rather than the fat per-token margin a closed API earns. The analog the forecast reached for is Hugging Face at roughly $100M ARR, not Anthropic at $1B in year three. The upside tail is a business-model shift, using the gigawatt of Vera Rubin capacity to sell first-party inference at scale. That is the path where a Fireworks-style $1B number opens up.
On the $50 billion round, 24%. The talks Bloomberg reported in November never closed, and Forge marks secondary trades around $13.5B today. There is also a specific gotcha holding the number down. The Nvidia and Google deals are exactly the kind of undisclosed-terms strategic money that never produces a public post-money print, so even a huge raise might not resolve this question YES. TML clearly needs the capital, since the research pegs the Nvidia capacity alone at something like $50B in compute costs, so a big financing event is coming. Whether it arrives with a five-handle valuation publicly attached is the 24%.
The frontier bet stays deferred. I give 12% to any TML model appearing in the Artificial Analysis top 10 before 2028. The mechanics are unforgiving. The top-10 threshold sits around 56 today and will drift into the 60s, Inkling scores about 41, and the board is crowded by multiple reasoning-effort variants of the same few frontier families, so a challenger has to reach frontier quality, not merely beat other open models. The gigawatt of Vera Rubin capacity only starts arriving in early 2027, which puts the first model trained on it right at the edge of this window. And TML has said plainly that Inkling is not trying to be the strongest model. That is the product-first strategy working as designed, and it is why the bench put TML's first top-5 model at a December 2030 median.
On the bench I wrote that Thinking Machines is the lab most likely to become a great business without ever holding the frontier. These three forecasts are that sentence turned into numbers, and each one resolves by the end of 2027. If Tinker's revenue comes in far above my range, the product-first path is vindicated faster than I expect. If the $50B round closes this year, the market is pricing the frontier bet, not the business. I will grade all three either way.
Forecast these yourself in the FutureSearch app , the moment a round closes, or the moment the next Inkling ships.